WO2023032360A1 - 画像処理装置、画像処理方法、および画像処理プログラム - Google Patents
画像処理装置、画像処理方法、および画像処理プログラム Download PDFInfo
- Publication number
- WO2023032360A1 WO2023032360A1 PCT/JP2022/021174 JP2022021174W WO2023032360A1 WO 2023032360 A1 WO2023032360 A1 WO 2023032360A1 JP 2022021174 W JP2022021174 W JP 2022021174W WO 2023032360 A1 WO2023032360 A1 WO 2023032360A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- class
- additional
- base
- neural network
- base class
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Definitions
- the present invention relates to image processing technology based on machine learning.
- CNN Convolutional Neural Network
- Continuous learning (incremental learning or continual learning) has been proposed as a method to avoid fatal forgetting.
- Continuous learning is a learning method in which when a new task or new data occurs, the model is not learned from the beginning, but the currently trained model is improved and learned.
- regularization-based continuous learning in which learning is performed using regularization loss (Patent Document 1).
- Patent Document 1 The technology described in Patent Document 1 has the problem that fatal forgetting cannot be sufficiently reduced.
- the present invention was made in view of this situation, and its purpose is to provide an image processing technology based on machine learning that can reduce fatal forgetting.
- an image processing apparatus provides input data based on an embedding vector output by a basic neural network that has already learned a base class and a barycentric vector of the base class.
- a base class selection unit that selects a base class by using a base class
- a continuous learning unit that continuously learns an additional class using an additional neural network that has already learned the base class
- an additional class selection unit that selects an additional class based on the embedding vector output by and the centroid vector of the base class and the additional class
- the base class selected by the base class selection unit and a classification determination unit that classifies the input data based on the added class.
- Another aspect of this embodiment is an image processing method.
- FIG. 1 is a configuration diagram of an image processing apparatus according to an embodiment
- FIG. 2 is a flowchart for explaining continuous learning processing by the image processing apparatus of FIG. 1
- 2 is a diagram illustrating the structure of a neural network model used in the basic neural network processing unit and the additional neural network processing unit in FIG. 1
- FIG. 2 is a flowchart for explaining classification determination processing by the image processing apparatus of FIG. 1;
- FIG. 1 is a configuration diagram of an image processing device 100 according to the embodiment.
- the image processing apparatus 100 includes a basic neural network processing unit 10, a basic class selection unit 20, an additional neural network processing unit 30, an additional class selection unit 40, a continuous learning unit 50, a centroid derivation unit 60, a centroid vector correction unit 70, and a classification A decision unit 80 is included.
- machine learning is performed by combining continuous learning and metric learning.
- an image will be described as an example of input data, but input data is not limited to an image.
- Metric learning is a technique for learning an embedding space (feature space) that considers the relationship between images (see, for example, Non-Patent Document 1).
- Metric learning is used in various fields such as information retrieval, data classification, and image recognition. Continuous learning using regularization loss can be combined with metric learning using metric loss.
- class incremental learning which is one type of continuous learning
- Non-Patent Document 2 performs class incremental learning with one neural network
- Non-Patent Document 3 performs class incremental learning using two neural networks with different learning rates, and performs classification using a combined feature space that combines the feature spaces of the two neural networks.
- the basic neural network that has learned the basic class is not changed, but the additional neural network that has already learned the basic class and continuously learns the additional class is updated.
- Class classification class selection
- the input image is classified into the class with higher accuracy (shorter distance).
- FIG. 2 is a flowchart explaining continuous learning by the image processing device 100.
- FIG. The configuration and overall operation of continuous learning will be described with reference to FIGS. 1 and 2.
- a neural network trained on base classes may be obtained from the network or trained using a dataset containing the base classes. It is desirable that the neural network that has already learned the base class has undergone metric learning (embedding learning) rather than class classification learning.
- the centroid vector of the base class may be obtained from the network, the image of the base class is input to the trained neural network, the centroid of the embedding vector output from the trained neural network is obtained for each class, and the may be derived as the centroid vector of
- the number of center-of-gravity vectors for each class is 1, but it may be more than one.
- a neural network that has learned the basic class is set in the basic neural network processing unit 10 and the additional neural network processing unit 30 (S10).
- the center-of-gravity vector of the base class derived using the neural network that has learned the base class is set in the base class selection unit 20 and the additional class selection unit 40 (S20).
- the base class selection unit 20 and the additional class selection unit 40 each store the centroid vector of the base class.
- the additional neural network processing unit 30 inputs all images of a certain additional class to the additional neural network before performing the learning session i, and Derive embedding vectors for all images.
- the centroid derivation unit 60 derives the centroid vector of the additional class from the embedding vectors of all the images of the additional class (S40).
- the center-of-gravity vector of the additional class here is the center-of-gravity vector before learning. Note that the center-of-gravity vectors of the additional classes are derived for all the additional classes.
- the continuous learning unit 50 continuously learns the additional neural network using the additional training data set including the additional class as learning session i (S50).
- the additional neural network processing unit 30 inputs all images of an additional class to the additional neural network after the learning session i, and Derive the embedding vector for all images of the class.
- the centroid deriving unit 60 derives the centroid vector of the additional class from the embedding vectors of all the images of the additional class (S60).
- the center-of-gravity vector of the additional class here is the center-of-gravity vector after learning. Note that the center-of-gravity vectors of the additional classes are derived for all the additional classes.
- the additional class selection unit 40 deletes the stored centroid vector of the base class (S70).
- the number of base class centroid vectors to be deleted is the number of additional classes added in learning session i.
- the centroid vector of the base class to be deleted is the closest neighbor to the centroid vector of the additional class added in learning session i.
- the centroid vector is not deleted.
- the number of center-of-gravity vectors stored in the base class selection unit 20 and the number of center-of-gravity vectors stored in the additional class selection unit 40 can be made the same.
- the center-of-gravity vector correction unit 70 corrects the center-of-gravity vector of the known class stored in the additional class selection unit 40 (S80).
- the known classes include the base class and the additional classes of learning session (i-1). Additional classes in learning session i do not need to be corrected.
- Non-Patent Document 2 For correction of the center-of-gravity vector of the learned (known) class, the method described with reference to FIG. 3 in Non-Patent Document 2 is improved and used.
- the center-of-gravity vector correction unit 70 corrects the learned class based on the center-of-gravity vector of the class before continuous learning and the center-of-gravity vector of the class after continuous learning within a predetermined distance from the center of gravity vector of the learned class (known class). Correct the centroid vector of . Specifically, the center-of-gravity vector correction unit 70 obtains the amount of movement of the center-of-gravity vector of the class after continuous learning from the center-of-gravity vector of the class before continuous learning, and calculates the average movement amount of these movement amounts. The center-of-gravity vector correction unit 70 corrects the center-of-gravity vector of the learned class by adding the average movement amount to the center-of-gravity vector of the learned class.
- Non-Patent Document 2 the correction is performed using the pre-learning embedding vector within the radius R of the center-of-gravity vector of the known class. The difference is that both of the centroid vectors are used for correction. In calculating the average amount of movement, using more centroid vectors is less likely to be affected by fine variations in each image. We decided to correct using both the center of gravity vector of the class before continuous learning and the center of gravity vector of the class after continuous learning.
- the basic training dataset is a supervised dataset that includes a large number of basic classes (eg, about 100 to 1000 classes) and each class consists of a large number of images (eg, 3000 images).
- the basic training dataset is assumed to be a sufficient amount of data to train a general classification task by itself.
- the additional training dataset is a supervised dataset containing a small number of additional classes (e.g., 2 to 10 classes), each of which consists of a small number of images (e.g., 1 to 5).
- a set of three images, an anchor image belonging to a certain class, a positive image belonging to the same class as the anchor image, and a negative image belonging to a class different from the anchor image, is input to the learning target neural network.
- the reason why the minority class is set to 2 is that even if the class to be learned is 1, it is necessary to have a class that is not to be learned as a negative image.
- the image is a minority image, but a large number of images may be used as long as the image is of a minority class.
- FIG. 3 is a diagram explaining the structure of a neural network model used in the basic neural network processing unit 10 and the additional neural network processing unit 30.
- FIG. A neural network is a deep neural network that contains convolutional and pooling layers and no fully connected layers. It includes ResNet-18 convolutional layers CONV-1 to CONV-5 shown in FIG. 3, followed by a global average pooling layer, which outputs a 512-dimensional embedding vector.
- the continuous learning unit 50 adds the metric loss Lml and the regularization loss Lr to calculate the total loss L as shown in the following equation, and learns the neural network so as to minimize the total loss L.
- L ⁇ (Lml+Lr)
- ⁇ indicates taking the sum for the input image.
- the regularization loss Lr is the embedding vector loss Lrv for minimizing the difference between the embedding vectors before and after the learning session, which is output when the image is input to the neural network as shown in the following equation.
- Lrv
- V(i-1) is the embedding vector output by the neural network for learning session (i-1).
- are symbols indicating the meaning of calculating the Frobenius norm.
- FIG. 4 is a flowchart explaining classification determination by the image processing apparatus 100.
- FIG. The configuration and overall operation of classification determination will be described with reference to FIGS. 1 and 4.
- the basic neural network processing unit 10 inputs the image to be classified into the basic neural network
- the additional neural network processing unit 30 inputs the image to be classified into the continuously learned additional neural network (S100).
- the basic neural network processing unit 10 supplies the embedding vector of the classification target image output from the basic neural network to the basic class selection unit 20, and the additional neural network processing unit 30 supplies the classification target image output from the additional neural network.
- the embedding vector of the image is supplied to the additional class selection unit 40 (S110).
- the base class selection unit 20 selects a base class based on the base embedding vector output by the base neural network (S120). Specifically, the base class having the centroid vector closest to the base embedding vector is selected.
- the additional class selection unit 40 selects an additional class based on the additional embedding vector output by the additional neural network (S130). Specifically, the additional class that has the centroid vector closest to the additional embedding vector is selected. Note that the additional class selection unit 40 does not select the base class even if the class having the centroid vector closest to the additional embedding vector is the base class.
- the classification determination unit 80 compares the base class selected by the base class selection unit 20 and the additional class selected by the additional class selection unit 40, and selects the class with the shorter distance between the center of gravity vector and the embedding vector. , is determined as the class of the classification result of the image to be classified (S140).
- the distance between the centroid vector and the embedding vector may be treated as the reciprocal of the probability, the degree of probability may be determined, and the class with the higher probability may be determined as the class of the classification result.
- the additional class is selected as the class of the classification result.
- the additional class selection unit 40 selects the centroid vector closest to the additional embedding vector regardless of whether it is the base class or the additional class.
- the classification determination unit 80 selects the base class selected by the base class selection unit 20 as the classification result. select as a class.
- the reason why the classification determination unit 80 selects the base class selected by the base class selection unit 20 as the class of the classification result is that the base neural network learns the base class with more data. . In other words, the classification determination unit 80 selects the classification result of the neural network that has learned with more data.
- the various processes of the image processing apparatus 100 described above can of course be implemented as an apparatus using hardware such as a CPU and memory, and can also be stored in a ROM (Read Only Memory), flash memory, or the like. It can also be realized by software such as firmware or software such as a computer.
- the firmware program or software program may be recorded on a computer-readable recording medium and provided, transmitted to or received from a server via a wired or wireless network, or transmitted or received as data broadcasting of terrestrial or satellite digital broadcasting. is also possible.
- the basic neural network since the basic neural network does not continuously learn, the basic class is not forgotten. Therefore, the basic neural network can classify the basic class with high probability even as the learning session progresses. Since the basic neural network does not continuously learn additional classes, the basic neural network cannot select additional classes. You can consider studying and choose additional classes.
- the classification result by the basic neural network that does not forget the basic class and the classification result by the additional neural network that continuously learns the additional class are evaluated, and the classification result with the higher accuracy is selected. Therefore, it is possible to improve classification accuracy while reducing fatal forgetting.
- the additional neural network learns only the additional classes, the number of data in the additional classes is small, so there is a high possibility that the centroid vectors of the additional classes will be overfitted. Also, the correction of the center of gravity vector is likely to be excessively corrected as well. Therefore, in the training of the additional neural network, by considering the embedding vector output by the basic neural network that is learning the basic class with a large amount of data together with the additional class, the centroid vector of the additional class and the correction of the centroid vector are overfitted. It prevents large fluctuations, and overfitting for additional classes of centroid vectors and centroid vector corrections is greatly reduced.
- the embedding space of the base neural network and the additional neural network can be maintained at the same density, and the base class can be selected.
- the embedding space distances in the unit 20 and the additional class selection unit 40 can be handled to the same degree. It is possible to prevent bias in class selection between the base class selection unit 20 and the additional class selection unit 40.
- the present invention can be used for image processing technology based on machine learning.
- 10 basic neural network processing unit 20 basic class selection unit, 30 additional neural network processing unit, 40 additional class selection unit, 50 continuous learning unit, 60 centroid derivation unit, 70 centroid vector correction unit, 80 classification determination unit, 100 image processing Device.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Image Analysis (AREA)
- Editing Of Facsimile Originals (AREA)
- Image Processing (AREA)
Abstract
Description
L=Σ(Lml+Lr)
ここで、Σは入力画像に対して和を取ることを示す。
Lml=dp-dn+α
ここで、dpは、アンカー画像の埋め込みベクトルとポジティブ画像間の埋め込みベクトルのユークリッド距離である。dnは、アンカー画像の埋め込みベクトルとネガティブ画像間の埋め込みベクトルのユークリッド距離である。αはオフセットである。
Lrv=||V(i)-V(i―1)||
ここで、V(i)は、学習セッションiのニューラルネットワークの出力する埋め込みベクトルである。V(i―1)は、学習セッション(i-1)のニューラルネットワークの出力する埋め込みベクトルである。||・||は、フロベニウスノルムを算出する意味を示す記号である。
追加クラス選択部40と分類決定部80の変形例について説明する。実施の形態とは異なる動作のみを説明する。追加クラス選択部40は、追加埋め込みベクトルと最も距離が近い重心ベクトルを、基本クラスであるか追加クラスであるかに関わらず選択する。ここで、基本クラス選択部20で選択された基本クラスと追加クラス選択部40で選択された基本クラスが異なる場合、分類決定部80は基本クラス選択部20で選択された基本クラスを分類結果のクラスとして選択する。ここで、分類決定部80が基本クラス選択部20で選択された基本クラスを分類結果のクラスとして選択する理由は、基本クラスについて基本ニューラルネットワークの方がより多いデータで学習しているからである。つまり、分類決定部80は、より多くのデータで学習しているニューラルネットワークの分類結果を選択するようにする。
Claims (6)
- 入力データに対して、基本クラスを学習済みの基本ニューラルネットワークが出力する埋め込みベクトルと、基本クラスの重心ベクトルとに基づいて基本クラスを選択する基本クラス選択部と、
基本クラスを学習済みの追加ニューラルネットワークを用いて追加クラスを継続学習する継続学習部と、
前記入力データに対して、継続学習された前記追加ニューラルネットワークが出力する埋め込みベクトルと、基本クラスおよび追加クラスの重心ベクトルとに基づいて追加クラスを選択する追加クラス選択部と、
前記基本クラス選択部により選択された基本クラスと、前記追加クラス選択部により選択された追加クラスとに基づいて、前記入力データをクラス分類する分類決定部とを備えることを特徴とする画像処理装置。 - 前記追加ニューラルネットワークが出力する埋め込みベクトルから重心ベクトルを導出する重心導出部と、
前記重心導出部により導出された継続学習前の重心ベクトルと継続学習後の重心ベクトルとに基づいて、継続学習前に既知のクラスの重心ベクトルを補正する重心補正部とをさらに備えることを特徴とする請求項1に記載の画像処理装置。 - 前記追加クラス選択部は、継続学習時の追加クラスの数だけ基本クラスの重心ベクトルを削除することを特徴とする請求項1に記載の画像処理装置。
- 前記追加クラス選択部が選択したクラスが基本クラスであり、前記追加クラス選択部が選択した基本クラスと、前記基本クラス選択部が選択した基本クラスが異なる場合、前記分類決定部は前記基本クラス選択部が選択した基本クラスを分類結果とする請求項1に記載の画像処理装置。
- 入力データに対して、基本クラスを学習済みの基本ニューラルネットワークが出力する埋め込みベクトルと、基本クラスの重心ベクトルとに基づいて基本クラスを選択する基本クラス選択ステップと、
基本クラスを学習済みの追加ニューラルネットワークを用いて追加クラスを継続学習する継続学習ステップと、
前記入力データに対して、継続学習された前記追加ニューラルネットワークが出力する埋め込みベクトルと、基本クラスおよび追加クラスの重心ベクトルとに基づいて追加クラスを選択する追加クラス選択ステップと、
前記基本クラス選択ステップにより選択された基本クラスと、前記追加クラス選択ステップにより選択された追加クラスとに基づいて、前記入力データをクラス分類する分類決定ステップとを含むことを特徴とする画像処理方法。 - 入力データに対して、基本クラスを学習済みの基本ニューラルネットワークが出力する埋め込みベクトルと、基本クラスの重心ベクトルとに基づいて基本クラスを選択する基本クラス選択ステップと、
基本クラスを学習済みの追加ニューラルネットワークを用いて追加クラスを継続学習する継続学習ステップと、
前記入力データに対して、継続学習された前記追加ニューラルネットワークが出力する埋め込みベクトルと、基本クラスおよび追加クラスの重心ベクトルとに基づいて追加クラスを選択する追加クラス選択ステップと、
前記基本クラス選択ステップにより選択された基本クラスと、前記追加クラス選択ステップにより選択された追加クラスとに基づいて、前記入力データをクラス分類する分類決定ステップとをコンピュータに実行させることを特徴とする画像処理プログラム。
Priority Applications (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP22863950.6A EP4398186B1 (en) | 2021-08-31 | 2022-05-24 | Image processing device, image processing method, and image processing program |
| CN202280055588.3A CN117813633A (zh) | 2021-08-31 | 2022-05-24 | 图像处理装置、图像处理方法以及图像处理程序 |
| US18/588,056 US20240212323A1 (en) | 2021-08-31 | 2024-02-27 | Image processing apparatus, image processing method, and non-transitory computer-readable medium having image processing program |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2021140819A JP7582127B2 (ja) | 2021-08-31 | 2021-08-31 | 画像処理装置、画像処理方法、および画像処理プログラム |
| JP2021-140819 | 2021-08-31 |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| US18/588,056 Continuation US20240212323A1 (en) | 2021-08-31 | 2024-02-27 | Image processing apparatus, image processing method, and non-transitory computer-readable medium having image processing program |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2023032360A1 true WO2023032360A1 (ja) | 2023-03-09 |
Family
ID=85411159
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2022/021174 Ceased WO2023032360A1 (ja) | 2021-08-31 | 2022-05-24 | 画像処理装置、画像処理方法、および画像処理プログラム |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US20240212323A1 (ja) |
| EP (1) | EP4398186B1 (ja) |
| JP (1) | JP7582127B2 (ja) |
| CN (1) | CN117813633A (ja) |
| WO (1) | WO2023032360A1 (ja) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2025153770A (ja) * | 2024-03-29 | 2025-10-10 | クラシエ株式会社 | 健康管理支援システム、健康管理支援装置、健康管理支援方法、及びプログラム |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117056726B (zh) * | 2023-07-25 | 2025-10-10 | 清华大学 | 一种面向消息传递神经网络的Wasserstein质心匹配层方法及产品 |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210124993A1 (en) * | 2019-10-23 | 2021-04-29 | Adobe Inc. | Classifying digital images in few-shot tasks based on neural networks trained using manifold mixup regularization and self-supervision |
-
2021
- 2021-08-31 JP JP2021140819A patent/JP7582127B2/ja active Active
-
2022
- 2022-05-24 EP EP22863950.6A patent/EP4398186B1/en active Active
- 2022-05-24 WO PCT/JP2022/021174 patent/WO2023032360A1/ja not_active Ceased
- 2022-05-24 CN CN202280055588.3A patent/CN117813633A/zh active Pending
-
2024
- 2024-02-27 US US18/588,056 patent/US20240212323A1/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210124993A1 (en) * | 2019-10-23 | 2021-04-29 | Adobe Inc. | Classifying digital images in few-shot tasks based on neural networks trained using manifold mixup regularization and self-supervision |
Non-Patent Citations (4)
| Title |
|---|
| HANBIN ZHAOYONGJIAN FUMINTONG KANGQI TIANFEI WUXI LI: "MgSvF: Multi-Grained Slow vs. Fast Framework for Few-Shot Class-Incremental Learning", ARXIV, vol. 2006, 2021, pages 15524 |
| LU YUBARTLOMIEJ TWARDOWSKIXIALEI LIULUIS HERRANZKAI WANGYONGMEI CHENGSHANGLING JUIJOOST VAN DE WEIJER: "Semantic Drift Compensation for Class-Incremental Learning", COMPUTER VISION AND PATTERN RECOGNITION, 2020, pages 6982 - 6991 |
| See also references of EP4398186A4 |
| THOMAS MENSINKJAKOB VERBEEKFLORENT PERRONNINGABRIELA CSURKA: "Distance-Based Image Classification: Generalizing to new classes at near-zero cost", IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, vol. 35, no. 11, 2013, pages 2624 - 2637, XP011527040, DOI: 10.1109/TPAMI.2013.83 |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2025153770A (ja) * | 2024-03-29 | 2025-10-10 | クラシエ株式会社 | 健康管理支援システム、健康管理支援装置、健康管理支援方法、及びプログラム |
| JP7812524B2 (ja) | 2024-03-29 | 2026-02-10 | クラシエ株式会社 | 健康管理支援システム、健康管理支援装置、健康管理支援方法、及びプログラム |
Also Published As
| Publication number | Publication date |
|---|---|
| JP2023034530A (ja) | 2023-03-13 |
| EP4398186B1 (en) | 2025-12-03 |
| JP7582127B2 (ja) | 2024-11-13 |
| EP4398186A4 (en) | 2024-12-25 |
| US20240212323A1 (en) | 2024-06-27 |
| EP4398186A1 (en) | 2024-07-10 |
| CN117813633A (zh) | 2024-04-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114387486B (zh) | 基于持续学习的图像分类方法以及装置 | |
| Yoon et al. | Lifelong learning with dynamically expandable networks | |
| US11562250B2 (en) | Information processing apparatus and method | |
| US20240212323A1 (en) | Image processing apparatus, image processing method, and non-transitory computer-readable medium having image processing program | |
| US12518168B2 (en) | Training and application method apparatus system and storage medium of neural network model | |
| CN109886343B (zh) | 图像分类方法及装置、设备、存储介质 | |
| WO2019045802A1 (en) | LEARNING DISTANCE MEASUREMENT USING PROXY MEMBERS | |
| CN112446487A (zh) | 神经网络模型的训练和应用方法、装置、系统及存储介质 | |
| US20220044125A1 (en) | Training in neural networks | |
| CN108710948A (zh) | 一种基于聚类均衡和权重矩阵优化的迁移学习方法 | |
| US11676027B2 (en) | Classification using hyper-opinions | |
| Heim et al. | A learnable safety measure | |
| WO2024024217A1 (ja) | 機械学習装置、機械学習方法、および機械学習プログラム | |
| US20240265257A1 (en) | Machine learning device, machine learning method, and non-transitory computer-readable recording medium embodied thereon machine learning program | |
| Diamant et al. | De-confusing pseudo-labels in source-free domain adaptation | |
| JP7350587B2 (ja) | 能動学習装置、能動学習方法及びプログラム | |
| CN111062406B (zh) | 一种面向异构领域适应的半监督最优传输方法 | |
| JP2021047797A (ja) | 機械学習装置、機械学習方法、及びプログラム | |
| JP2024173962A (ja) | 機械学習装置、推論装置、機械学習方法、および機械学習プログラム | |
| CN118876073A (zh) | 基于分布式鲁棒元强化学习的机器人运动控制方法 | |
| CN116776119A (zh) | 一种抑制灾难性遗忘的方法、装置及相关介质 | |
| Hofer et al. | Research Agenda for Reducing Feature Descriptor Sizes in Networked Visual-SLAM | |
| US20260011138A1 (en) | Image classification apparatus, image classification method, and non-transitory computer-readable medium having image classification program | |
| JP7735828B2 (ja) | 機械学習装置、機械学習方法、および機械学習プログラム | |
| Gogoi et al. | Perturbing the gradient for alleviating meta overfitting |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22863950 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202280055588.3 Country of ref document: CN |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2022863950 Country of ref document: EP |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| ENP | Entry into the national phase |
Ref document number: 2022863950 Country of ref document: EP Effective date: 20240402 |
|
| WWG | Wipo information: grant in national office |
Ref document number: 2022863950 Country of ref document: EP |